Comparable Corpora BootCaT (CCBC)

Adam Kilgarriff, Avinesh PVS, Jan Pomikalek

Lexical Computing Ltd.

Just-in-time corpora

Krista Varantola

Translators, terminologists

In-domain terminology: Domain dictionaries

• Don’t exist

• Out of date

• Not accessible

Collect in-domain web pages

Instant corpus

2

BootCaT (Bootstrapping Corpora and Terms)

Baroni and Bernardini 2004

User: input ‘seed terms’

Send 3-at-a-time to a search engine• Returns search hits page

Retrieve those pages

A corpus!• Cleaning, deduplicating, linguistic processing

Extract terms• Can use extracted terms as seeds, iterate

3

Works well

Widely used More implementations

SkE has WebBootCaT, web front end Secret:

piggybacks on search enginesThey do the donkey-work

• on-domain, text-rich pages, no spam, …

4

Also in use for

General language corpusLong list of general seed words

• Pioneer: Serge Sharoff• LCL: Corpus Factory

‘Varieties of Learner English’General English, same queries except

• Region=UK, US, Canada, Aus, China, Japan, Korea

Validation under way

5

Corpus query tool, since 2003

Widely used by lexicographers• Commercial

• OUP, CUP, Collins, Macmillan, Le Robert, Cornelsen, Shogukakan

• National dictionary projects• Bulgaria, Czech Republic, Estonia, Netherlands, Slovakia,

Slovenia

Universities• Linguistics, language research, NLP, language

teaching, teaching translation

6

The Sketch Engine

55 languages and counting

Large corpora ready-to-use for

Arabic Bengali Bulgarian Chinese Czech Croatian Danish Dutch English Estonian Finnish French German Greek Gujarati Hebrew Hindi Indonesian Irish Italian Japanese Korean Latin Malay Malayalam Norwegian Persian Polish Portuguese Romanian Russian Serbian Setswana Slovak Slovene Spanish Swahili Swedish Tamil Telugu Thai Turkish Urdu Vietnamese

7

Handles large corporaLargest to date: 8 billion words

Fast Web-based: no software to install Build ‘instant corpora’ from the web Load your own corpus

Quota of space on SkE server Word sketches

One-page, automatic accounts of a word’s grammatical and collocational behaviour

Free 30-day trial: sketchengine.co.uk8

9

Adam Kilgarriff

Lexical Computing Ltd.

WebBootCaT

BootCaT integrated in SkE BootCaT a corpus

Clean, de-dupe, POS-tag, thenLoad into Sketch Engine

10

How big a corpus do we get?

Observation

Specialist domain, L1 Specialist domain, L2 Matching terminology

16

Going multilingual

Translate seeds English: volcanology volcanologist "volcanic

eruption" seismographs Eyjafjallajokull geodic "deformation monitoring" tephra magma stratigraphic tephrochronology geochronological "volcanic ash" ablation rhyolitic

French:vulcanologue volcanologie "éruption volcanique" sismographes Eyjafjallajokull "surveillance de la déformation" géodiques tephra magma téphrochronologiestratigraphique géochronologiques "de cendres volcaniques" ablation rhyolitiques

BootCaT for French

CCBC

Input: L1, L1 seeds, L2 Choose dictionary

Google as default• Google dictionary (25 lg pairs, limited API)• Google translate (1225 lg pairs, only 1 transl)

Option: edit translations Bootcat 2 corpora Bilingual word sketches

19

Bilingual word sketches(very first pass)

For L1 nodeword nFor each of its translations n1, n2, …

• For each collocate c in word sketch• For each of its translations c1, c2, …

• Does ci occur as collocate in word sketch for ni?

• If yes: output <c; ni , ci >

• Add L1 and L2 examples sentences

20

21

Matching seeds – how?

User translates Yes but limited

Bilingual dictionary Yes but finding them?? Google dictionary

Machine translations Wikipedia

Matching articles

Evaluation

Extract terms for L1, L2 Ask expert

1. Are they terms

2. Do the L1, L2 lists contain translations of each other?

23

3 lg-pairsEn-Fr, En-De, En-CzOne expert for each pair

3 domainsVolcanoesStradivariusPancreatic cancer

• Wikipedia: En and De only

24

Results

25

In brief

•Words good•Multiwords bad

Unithood and termhood

To find termsFor multiwords only

• Does it hang together?

• UnithoodIt it distinctive?

• Keywords

• Termhood

We didn’t use termhood for multiwords but need to

26

Next steps

Termhood for multiwords WebBootCaT from wikipedia From collocations to terms

More-than-2-word collocations• … deadline next Tuesday

27

Thank you

http://www.sketchengine.co.uk

28